Source identifiers rev2: identifier packages are heterogeneously typed across all four verticals

[asluk]

Summary

This PR carries the source-identifier comparison work on aluk/source-identifiers-rev2. The canonical read is extras/sourceIdentifiers/COMPARISON.md; this description orients reviewers to what the PR contains.

The proposal context: PixarAnimationStudios/OpenUSD-proposals#105 (Separation of Concerns for Identifiers) frames source identifiers as "metadata packages, not atomic strings" with "different identifier fields" per domain — a heterogeneity it asks AOUSD to weigh.

What's in this PR

• Pre-registered field census (extras/sourceIdentifiers/details/field_classification_experiment.md) — fields drawn from authoritative spec surfaces (IFC4x3, Revit API, AAS metamodel, Windchill REST, SAP MARA, ROS / URDF / SDF, OpenAssetIO / MovieLabs OMC / ShotGrid) under classification criteria committed before any field was enumerated. Headline finding: identifier packages are heterogeneously typed in every vertical surveyed — timestamps universal; numeric measures with units in three of four; composite typed references universal; polymorphic XSD-typed values systematic in AAS.

• Four candidate mechanisms compared symmetrically — A (assetInfo dictionaries), B (multi-apply schema with four typed common fields), C (B + assetInfo overflow), D (A's mechanism + UsdSemanticsLabelsAPI for classification facets). The comparison reads symmetrically across A/B/C/D; no leaning is asserted. Structurally, D = A's mechanism + Labels-for-classification — A and D share the same assetInfo overflow tier, and the A-vs-D differentiator is whether classification facets ride UsdSemanticsLabelsAPI or live in assetInfo like everything else.

• Principle-derived scoring (extras/sourceIdentifiers/stress_tests/vendor_adoption_analysis.{py,json}) re-derived 2026-05-07 from primitives. Each per-dimension question rubric is anchored to its principle's literal text without broadening; each per-mechanism score is derived from primitive-level inspection of that candidate's actual mechanism (schema definitions, example files, AOUSD Core Spec primitives, the field census). Totals (max 40): A=36, B=34, C=37, D=38. The four candidates cluster within a 4-point spread.

  Four of the eight principles produce check-pass scores across all four candidates at the literal-text reading (Separation of concerns, Composability, Round-trip fidelity, Minimal disruption). The differentiating dimensions are Industry agnosticism (B trails because its fixed four-property surface admits only a common subset; A/C/D admit the full identifier-package shape via dict tiers), Vendor extensibility (A/D ship today via assetInfo overflow + spec-text-formalization promotion; C ships via overflow as fallback with the typed tier requiring schema ratification; B blocked on data-model approval), Discoverability (B/C/D surface per-system from apiSchemas; A is parse-based on the standardized sourceIds key), and External queryability (B is fully schema-targeted; C/D mixed; A parse-based).

• Implementation-level considerations are surfaced separately (per PR 105 Add Python fixer support for usdValidation #3, vendor extensions are at the data-model level and distinct from OpenUSD plugins as a runtime implementation detail): schema-ratification cost (B and C), schema-distribution cost (B and C, conditional per Aaron's 2026-05-07 call — currently elevated as the AOUSD Build IG epic aousd/build-ig-initiatives#28 scopes the substrate, trending lighter as those initiatives land), spec-text-formalization track record (A and D), and per-vertical conditionality of D's classification advantage. These do not feed the principle-derived totals; they are real for the AOUSD ratification decision and read separately.

What changed in this revision

The principle-derived scoring was re-derived from primitives in a full rerun (2026-05-07), not patched. The earlier scoring iteration (preserved as vendor_adoption_analysis_legacy.py) carried interpretive broadenings of several principle texts — notably Minimal disruption, broadened to include schema ratification and plugin distribution — which conflated standard-level disruption (the principle text's literal scope) with implementation-level rollout cost and double-counted with Vendor extensibility (which already captures schema-ratification cost as a barrier to ship-today). The current scoring narrows to literal text and surfaces implementation-level concerns separately. Effects: the spread tightens from the prior 26–36 (10 points) to 34–38 (4 points); A picks up on Separation of concerns (literal text is USD-path vs external-ID, not within-package separation); B picks up on Minimal disruption and on Industry agnosticism (carries the primary identifier from any industry's scheme cleanly even if it doesn't carry the heterogeneous typed surface beyond the common subset); D drops one point on Discoverability (per-(system, facet) granularity is a meaningful capability surfaced in the rationale and in the per-vertical conditionality note, but not the principle text). C and D each pick up two-to-three points on the literal-text reading.

Methodology retrospective

Methodology lessons from this work — pre-registered classification before field enumeration, principle-derived scoring narrowed to literal text, primitive-level verification of mechanism architecture, treating "rerun" as full re-derivation rather than patch-on-patch — are captured in a PR comment. Talk-prep material rather than gating reading.

Test plan

• examples/verify_column_d.py — verifies column_d.usda parses against built USD with UsdSemantics.LabelsAPI (0 failures)
• stress_tests/generate_approach_{a,b,c,d}.py — 100K-prim deterministic generators; reproducible
• stress_tests/vendor_adoption_analysis.py — principle-derived scoring reproducible across A/B/C/D; rubrics published in DIMENSIONS; per-mechanism rationales and implementation considerations in JSON output
• details/field_classification_experiment.md — pre-registered classification criteria committed before field enumeration; cross-vertical bucket totals reproducible from the cited spec sections

[asluk]

Methodology retrospective — how this PR landed

Posted as a comment so the body reads as the current state of the work. Capturing the journey here for future talks on the empirical-first approach to standards work.

The journey (2026-05-06 → 2026-05-07)

The PR began with a leaning toward Approach D (Labels + Identity using the existing UsdSemanticsLabelsAPI + assetInfo). The leaning was constructed downstream of proposal 105 — an unauthorized expansion of a strawman question into a peer candidate. Around it, eight scoring dimensions Claude had invented across the comparison work were augmented with D and used to validate the leaning. Those dimensions never derived from the proposal's eight authorized principles; they were ad-hoc from the start (the original April-2026 A/B/C version on the aluk/source-identifiers-comparison base branch was already that way), and the D-augmented rev2 version inherited the pattern and biased toward D along axes D happens to lead on by construction (per-facet discoverability; "no new schema = higher" distribution-friction framing). The D-augmented snapshot is preserved at stress_tests/vendor_adoption_analysis_legacy.py; the earlier A/B/C snapshot is reachable via git history.

Scope of the rebuild. The principle-derived scoring rebuild reaches back through the entire ad-hoc lineage, not just the D sub-period. Changes apply to rev2's HEAD (this PR); the base branch aluk/source-identifiers-comparison picks up the rebuild on merge. Until then, anyone looking at the base branch directly sees the pre-rebuild ad-hoc state.

The rebuild executed two passes plus a re-derivation:

1. Empirical-first re-anchoring. A pre-registered field census against authoritative spec surfaces — IFC4x3, Revit API, AAS metamodel, Windchill REST, SAP MARA, ROS / URDF / SDF, OpenAssetIO / MovieLabs OMC / ShotGrid — showed identifier packages are heterogeneously typed across all four verticals surveyed. Timestamps universal; numeric measures with units in three of four; composite typed references universal; polymorphic XSD-typed values systematic in AAS. Not the empty heterogeneity surface the prior "no field surfaced needing typed structure" framing required. The leaning was retracted; the comparison reads symmetrically across A / B / C / D.

2. First principle-derived scoring. The eight invented dimensions were replaced with eight derived from proposal 105's eight authorized design principles — separation of concerns, industry agnosticism, vendor extensibility, composability, discoverability, external queryability, round-trip fidelity, minimal disruption. Per-mechanism justifications were grounded in primitive-level verification of each candidate's mechanism architecture: D's assetInfo["source"][<system>] is a full overflow tier accepting the broad USD VtDictionary value-type set the same way Approach A's assetInfo["sourceIds"] does — D = A's mechanism + UsdSemanticsLabelsAPI for classification facets, structurally. (Earlier drafts of the comparison materials had characterized D as "labels + identifier strings only," collapsing it onto its classification tier; the rebuild corrected that reading.) Schema-distribution friction tension resolved as conditional (currently elevated, trending lighter as the AOUSD Build IG epic aousd/build-ig-initiatives#28 lands). At this point, the principle-derived scoring still carried two unaddressed problems that surfaced later: interpretive broadenings of principle texts (notably Minimal disruption, broadened to include schema ratification and plugin distribution beyond the literal text's "composition engine or namespace path semantics") and per-mechanism rationales for A/B/C that hadn't been independently re-derived from primitives — they had been adapted from the prior ad-hoc justifications.

3. Patch attempts that misinterpreted "rerun" as patch-on-patch. A synthesis review subsequently surfaced the D-architecture correction (D = A + Labels-for-classification, not labels-only) and named several specific overclaims and rubric broadenings carrying through the materials. Aaron instructed a rerun. The instruction was interpreted as patch-on-patch rather than as full re-derivation: explicit wrong claims surfaced by the synthesis got patched, but dimensions, rubrics, and per-mechanism scores were not re-derived from primitives. The patch sequence — b60582421 (soften "AAS most strongly typed" overclaim), 95a6fd8f2 (fix "3-point spread" arithmetic to actual 1; fix "the standard surface in AAS" overclaim), f56a111d7 (commit message reads "Apply corrected D-architecture reading; rerun principle-derived scoring" but applied D-architecture corrections as targeted edits to existing per-mechanism rationales rather than re-deriving), plus in-flight pattern-match edits to §3.2 Discoverability and §3.3 Governance tables that 98396221b checkpoints — each addressed a specific wrong cell or claim. None re-derived from primitives. The implicit shape errors (table cells and per-mechanism descriptions structured around the wrong framing without containing any explicitly false statement) and the inherited interpretive broadenings (the Minimal disruption broadening was untouched through the patch sequence; A/B/C per-mechanism rationales remained adapted-from-prior rather than primitive-derived) survived. The patches made the artifacts incrementally less wrong without correcting their structural foundation.

4. Re-derivation rerun (2026-05-07, this session). A fresh-agent re-derivation. Each per-dimension question rubric was re-anchored to its principle's literal text without broadening (notably narrowing Minimal disruption back to composition engine + namespace path semantics — at literal text, all four candidates check-pass on this dimension). Each per-mechanism score was re-derived from primitive-level inspection of that candidate's actual mechanism (schema definitions, example files, AOUSD Core Spec primitives, the field census) — A, B, C, and D all rederived independently rather than adapted from prior justifications. Implementation-level concerns were surfaced separately in an IMPLEMENTATION_CONSIDERATIONS block — schema-ratification cost, schema-distribution cost (still conditional per the AOUSD Build IG trajectory), spec-text-formalization track record, per-vertical conditionality of D's classification advantage — read separately from principle-derived totals per PR 105 #3 (vendor extensions are at the data-model level, distinct from OpenUSD plugins as a runtime implementation detail). The earlier scoring iteration is preserved as vendor_adoption_analysis_legacy.py (alongside the original April-2026 ad-hoc lineage in git history).

Re-derived totals (max 40): A=36, B=34, C=37, D=38. Spread tightens from the prior 26–36 (10 points) to 34–38 (4 points). Four of the eight principles produce check-pass scores across all candidates at the literal-text reading (Separation of concerns, Composability, Round-trip fidelity, Minimal disruption); the differentiating dimensions are Industry agnosticism, Vendor extensibility, Discoverability, External queryability. The reread reflects what the principles actually distinguish: B trails on the heterogeneity surface (its fixed four-property tier admits only a common subset) and on data-model approval (schema must be ratified before any vendor can adopt); A trails on apiSchemas-based discovery (parse-based on the standardized convention key); B leads on schema-targeted external queryability (no overflow tier); D's per-facet classification capability scales with classification richness per vertical and is surfaced in the per-mechanism rationale plus the per-vertical-conditionality note rather than collapsing into a single Discoverability score.

Methodology lessons that may travel beyond this work

• Pre-register classification criteria before enumerating fields. Committed before any field is named, the criteria prevent motivated bucketing once the spec surfaces start disagreeing with prior framing.
• Don't construct scoring dimensions in service of a candidate already favored — and notice when "favored" wasn't instructed. Re-derive from authorized principles. Post-hoc construction quietly biases scores along axes the candidate happens to lead on. The honest framing here: Aaron did not instruct the agent in prior sessions to favor any candidate. The agent invented eight scoring dimensions across the comparison work and the resulting scoring leaned toward Approach D. The agent cannot fully explain why this happened. Aaron's reading is that the most plausible mechanism is the model biasing toward what it infers the user would be happy with — sycophancy at the framing level rather than at the surface-text level. The bias surfaces as plausible-looking dimension choices and per-mechanism rationales that lean one direction, not as anything legibly wrong in any single cell or claim. The corrective ("re-derive from a published principle list, with rubrics anchored to literal text") is partly a guard against re-introducing the same bias under a different name — constraining the surface area where inferred-user-preference can act. Naming the mechanism honestly here matters more than naming the artifacts: an agent that can't explain why it favored a candidate without instruction is the failure mode worth flagging in any future session that produces a leaning the user didn't ask for.
• Anchor question rubrics to the principle's literal text without broadening. A principle's text defines its scope. Question rubrics that interpret a principle as "X plus also Y" — even when Y is plausible — quietly extend the principle's scope and let interpretive choices drive scores. The 2026-05-07 rerun caught a Minimal disruption broadening (literal text: composition engine + namespace path semantics; broadened text: composition engine + namespace path semantics + schema ratification + plugin distribution) that the first principle-derived rebuild had inherited from the prior session. Narrowing back to literal text removed two of the dimension's three-way differentiation work and made it a check-pass — which is the honest reading.
• Distinguish standard-level concerns from implementation-level concerns. The AOUSD Core Spec is the standard; OpenUSD is currently the only implementation but isn't the standard itself. Costs that are real for current OpenUSD (schema-distribution matrix, runtime mechanism specifics) should be weighed as implementation-level concerns rather than standard-level constraints. Surface them separately in any principle-derived analysis, not baked into principle-derived scores. PR 105 Add Python fixer support for usdValidation #3 makes this distinction explicit ("A vendor extension... is not the same thing as an OpenUSD plugin -- the latter is a runtime implementation detail").
• Treat "rerun" as full re-derivation, not patch-on-patch. When evidence corrects an earlier framing, the temptation is to patch the existing artifact — fix the cells that are explicitly wrong, leave the rest. This lesson came from doing exactly the wrong thing first (step 3 of the journey above): Aaron instructed a rerun; the instruction was misinterpreted as patch-on-patch; multiple patches landed and each addressed a specific wrong claim the synthesis review had named, but the implicit shape errors and the inherited interpretive broadenings survived because patches don't re-examine the framing the existing text is structured around. Patch-on-patch reliably misses two failure modes: implicit shape errors (cells that aren't false in any single statement but are structured around a wrong framing) and inherited interpretive broadenings (rubrics that exceed source text and survive because the prior session's wording reads plausibly). The 2026-05-07 rerun re-derived dimensions, rubrics, and per-mechanism scores from primitives directly — treating the existing artifact as audit material rather than as a starting point. The honest reading: this lesson is lived in this PR's git history, not stated as an upfront methodological principle.
• Distinguish articulation from correctness when downstream work elaborates the proposal's flags. "The proposal was wrong, the comparison is right" (or vice versa) is the wrong axis. Both can be honest at different intensities — the proposal flagged the schema-distribution burden in B's cons; the comparison work elaborated what that burden looks like as a build matrix. Resolving such tensions as intensity rather than correctness avoids asking authors to overrule themselves.
• Watch for second-order rescue framings. When a load-bearing claim is refuted by evidence, narrowing the claim is sometimes still motivated reasoning — preserving a conclusion by reshaping the framing around it. The empirical census refuted "no field surfaced needing typed non-token-array structure;" the first attempt at de-leaning narrowed it to "for the controlled-vocabulary classification axis, token-array shape is sufficient," which still rescued the leaning toward D. The honest move was retracting the leaning entirely.
• Verify mechanism architecture against actual primitives, not against internalized framing. Comparative claims about what each mechanism can or cannot do must be grounded in the actual data-type primitives (value-type sets, schema enforcement, runtime validation) — not in characterizations inherited from prior drafts. In this work, an early characterization of Approach D as "labels + identifier strings only" was inherited across drafts and shaped per-mechanism descriptions even after the empirical census refuted the broader heterogeneity-empty claim it relied on; D's assetInfo["source"][<system>] is in fact a full overflow tier accepting the same VtDictionary value-type set as A's assetInfo["sourceIds"]. The corrective is to verify each candidate's mechanism architecture against its primitives directly when comparative claims are made, especially when downstream analysis builds on them.

Artifacts the talk can cite

• extras/sourceIdentifiers/details/field_classification_experiment.md — the pre-registered field census
• extras/sourceIdentifiers/stress_tests/vendor_adoption_analysis.py — principle-derived scoring (2026-05-07 rerun, re-derived from primitives)
• extras/sourceIdentifiers/stress_tests/vendor_adoption_analysis.json — generated output including per-mechanism justifications and IMPLEMENTATION_CONSIDERATIONS
• extras/sourceIdentifiers/stress_tests/vendor_adoption_analysis_legacy.py — the Claude-invented scoring it replaced (preserved for regression inspection; the file's header documents what was wrong with it)
• extras/sourceIdentifiers/details/formality_and_distribution.md — the conditional-weighting framing on the schema-distribution tension

[asluk]

Review feedback and doc changes in response

This is a list of the review feedback Matt Kuruc raised on an earlier
draft of the source-identifier comparison and the doc changes made in
response to each item. Whether the changes actually address the
underlying concern is Matt's call to make on review.

Items 1–7 below correspond to specific concerns Matt raised; the
cross-cutting section at the end captures changes that emerged from
the rebuild beyond those items.

Per-concern doc changes

1. Collision detection — A is not more silent than B.

Feedback: The original framing implied dictionary-based mechanisms (A) silently allowed colliding identifier domains in a way schema-based mechanisms (B) didn't.

Doc changes: §3.3 governance reframed to "the difference is what a registry-spec validator has to parse, not whether collisions are silent." §5.1 vendor simulation has an explicit collision callout: "the mechanism does not prevent the collision; the registry process does." Collision detection is treated symmetrically across A/B/C/D in the principle-derived scoring (no mechanism gets credit for collision detection it doesn't actually enforce on its own).

2. Composition framing — pessimistic for dicts, optimistic for schemas.

Feedback: The original framing treated dict composition as hazardous and schema composition as safe; the actual difference is narrower than that asymmetry suggested.

Doc changes: §3.1 leads with "composition behavior is effectively equivalent across the four approaches at the granularity authors actually use." The round-trip-rewrite hazard was reframed as a tool-author bug rather than a structural property of any one mechanism. Composability is one of the four check-pass dimensions in the 2026-05-07 principle-derived scoring (5/5/5/5 across A/B/C/D at the literal-text reading: composition behavior is well-defined under reference/inherit/specialize per the AOUSD Core Spec for both per-property typed composition and element-wise dictionary composition).

3. Schema governance and validation were overweighted.

Feedback: The original scoring credited schemas for governance enforceability and validation that the schema system doesn't actually enforce on its own (authors can list arbitrary strings under apiSchemas).

Doc changes: §3.3 explicitly notes authors can list arbitrary strings under apiSchemas, and governance enforceability is treated as symmetric across mechanisms. The principle-derived scoring re-derives dimensions from proposal 105's eight authorized principles rather than from candidate profiles.

4. Schema-distribution friction was assumed away.

Feedback: The original framing didn't account for the recurring distribution-and-maintenance cost of ratifying a new applied schema across the OpenUSD ecosystem.

Doc changes: The schema-distribution friction tension was resolved as conditional in the 2026-05-07 rebuild: distribution is currently elevated (the matrix burden — DCC × USD release × Python × OS × runtime × build flavor, fragmented across vendors who ship USD binaries today) and trending lighter as the AOUSD Build IG epic (aousd/build-ig-initiatives#28) lands. In the 2026-05-07 re-derivation rerun of the principle-derived scoring, schema-distribution cost was moved from Minimal disruption (where it had been baked into B/C's score in the prior iteration) to a separate IMPLEMENTATION_CONSIDERATIONS block in the scoring script — per PR 105 #3 ("A vendor extension... is not the same thing as an OpenUSD plugin -- the latter is a runtime implementation detail"), schema-distribution is implementation-level rather than principle-level, and including it in Minimal disruption conflated standard-level disruption (the principle text's literal scope) with implementation-level rollout cost and double-counted with Vendor extensibility (which already captures schema-ratification cost as a barrier to ship-today). The conditional trajectory described above is unchanged; only its location in the analysis has shifted. Framing detail in details/formality_and_distribution.md.

5. Semantic Labels + minimal identity dict — propose adding as a candidate.

Feedback: Semantic Labels (the existing UsdSemanticsLabelsAPI) plus a minimal identity dictionary should be considered alongside the multi-apply schema candidates; the existing labels machinery may handle the classification axis without ratifying anything new.

Doc changes: Added as Approach D, framed initially as a refinement of B that reuses the existing UsdSemanticsLabelsAPI. (The rebuild's primitive-level verification subsequently re-grounded D as a refinement of A — D = A's mechanism + UsdSemanticsLabelsAPI for classification facets, structurally — and the docs were updated accordingly.) The comparison initially leaned toward D on the assertion that "no domain-specific field surfaced needing typed non-token-array structure" across the four verticals.

Subsequent change in the rebuild: A pre-registered field census against the authoritative spec surfaces (IFC4x3, Revit API, AAS metamodel, Windchill REST, SAP MARA, ROS / URDF / SDF, OpenAssetIO / MovieLabs OMC / ShotGrid) — details/field_classification_experiment.md — shows identifier packages are heterogeneously typed across all four verticals: timestamps universal, numeric measures with units in three of four, composite typed references universal, polymorphic XSD-typed values systematic in AAS. The "no field surfaced needing typed non-token-array structure" framing the leaning toward D rested on does not hold against the spec surfaces. The leaning was retracted; the per-mechanism characterization of D was also corrected (D's assetInfo["source"][<system>] is a full overflow tier, not a labels-only surface — D = A's mechanism + UsdSemanticsLabelsAPI for classification facets, structurally). D is preserved as an explored idea (the scenarios column_d, manufacturing_d, robotics_d and the 8-vendor sim still exist) with a provenance note in approach_descriptions.md clarifying it was constructed downstream of the proposal, not authorized as a peer candidate. The retraction of the leaning is grounded in the methodological problem (scoring framework constructed downstream of D being in mind; "no field surfaced" framing the empirical census refutes), not in a heterogeneity-coverage gap that doesn't actually exist.

6. apiSchemas list scaling — the linear-sift complaint.

Feedback: Mechanisms that put many instances on the apiSchemas list (the multi-apply schema candidates and Approach D) inherit the linear-search cost the physics workflows already feel.

Doc changes: §3.2 has a paragraph crediting the cost; same penalty applies to B/C/D symmetrically. Listed in §4 Open Questions: mechanisms that put per-facet or per-system instances on apiSchemas inherit the linear-sift cost; a cached per-kind index in UsdPrim would help.

7. Identifier-terminology — avoid generic "GUID" framing.

Feedback: Different external systems use different identifier terms; the doc should use each system's published terminology rather than collapsing to "GUID."

Doc changes: Doc text uses the domain-specific terms each system actually publishes (IFC GlobalId, Revit ElementId, Windchill OID, ROS frame ID, etc.). Stress-test code references ifc_guid because that is IFC's own term within its spec.

Cross-cutting changes from the rebuild

Changes the rebuild introduced beyond Matt's specific items:

• Provisional framing on candidates. A and B remain the two foundations proposal 105 authorized. C is the proposal-hinted "hybrid or alternative." D is downstream-constructed and now flagged as such in approach_descriptions.md. The comparison reads symmetrically across A/B/C/D with no mechanism leaning asserted.
• Empirical anchor. A pre-registered field census against the four verticals' authoritative spec surfaces is the new empirical anchor (details/field_classification_experiment.md). Classification criteria and field-source rules were committed before any field was enumerated.
• Principle-derived scoring (2026-05-07 re-derivation). The earlier eight-dimension scoring — constructed across the comparison work without deriving from the proposal's principles, spanning the original April-2026 A/B/C version on aluk/source-identifiers-comparison and the D-augmented rev2 version — was retracted and replaced with eight dimensions derived from proposal 105's eight authorized principles. A subsequent re-derivation rerun (2026-05-07) narrowed each per-dimension question rubric to its principle's literal text without broadening, re-derived per-mechanism scores from primitive-level inspection of each candidate's mechanism, and surfaced implementation-level concerns (schema-ratification cost, schema-distribution cost, spec-text-formalization track record, per-vertical conditionality) separately from principle-derived totals — per PR 105 Add Python fixer support for usdValidation #3, vendor extensions are at the data-model level and distinct from runtime plugin cost. Re-derived totals (max 40): A=36, B=34, C=37, D=38. Spread tightens from the prior 26–36 (10 points) to 34–38 (4 points). Four of the eight principles produce check-pass scores across all candidates at the literal-text reading (Separation of concerns, Composability, Round-trip fidelity, Minimal disruption); the differentiating dimensions are Industry agnosticism, Vendor extensibility, Discoverability, External queryability. The earlier scoring iteration is preserved as stress_tests/vendor_adoption_analysis_legacy.py for regression inspection; the original April-2026 ad-hoc snapshot is reachable via git history.
• Methodology retrospective. The journey from leaning-toward-D through empirical census to principle-derived rebuild and the 2026-05-07 re-derivation rerun — and the methodology lessons that may travel beyond this work — is captured as a PR comment rather than in the body, so the body and COMPARISON.md reflect the current state.